Would a new X mode be prudent? Options

[mithrandir]

Am I right in thinking that LAME could use another noise measurement mode (-X)? It's not that the 8 modes it currently has isn't enough but I think we need another one to add flexibility in tuning. Mode 3 is generally the best overall for quality purposes yet it can lead to significant bitrate bloat. Would it be prudent to add a new mode so that the criterions are:

The actual is better if the total sum of noise is lower than the best so far
and
the maximum noise over all scalefactor bands is less than the best so far plus 4dB.

Basically, a "weaker" -X 3 that should cut down the bitrate yet have higher quality than -X 1. The 4dB is arbitrary...it's just higher than mode 3's 2dB threshold.

[Dibrom]

I don't think this would be a good solution.

The reason that it's hard to get LAME to sound good at lower bitrates isn't because of the X modes, but because the psymodel is not very well tuned for this in many other areas. I think a total redesign would be more effective.

When I was working on the --alt-presets I experimented a lot with trying to create a new noise measuring mode for vbr-old/cbr with better criteria, but wasn't very successful. I believe the reason for this is that the overall system is too flawed for such a simple change to have any significant impact. The only good solution I found to lower bitrates while still offering improved quality was to use a different and much more aggressive noise measuring criterion for short blocks.

Also, -X 3 is not without problems, especially at lower bitrates. This is why it was never the only mode used with the --alt-presets (and why, IIRC, only --alt-preset insane used -X 3 for both long and short blocks, since the bitrate was already forced so high). The problems I encountered manifested themselves mostly on classical type pieces where the background noise was very audible. In these situations, -X 3 at lower bitrates would cause dropouts throughout the entire upper frequency range.. sounding horrible. The only solution was to use -X 1 on long blocks (which would end up being used 95% of the time or more), and -X 3 only for short blocks. This seemed to me to be a much more acceptable solution and kind of the "best of both worlds".

Anyway.. I don't think that a new -X mode is really a good idea. Almost all except for a few of the -X modes are already worthless.. a new one won't offer much improvement IMO. What we need is a new psymodel or some rather large changes to the current one.

At any rate, I suggest that if you really think this new mode would make much of a difference (obviously I don't ), you could always just add it into the source and recompile. The change would be simple enough.

Oh, and I haven't kept up with the development of 3.94 for obvious reasons... so it's possible that some things may have changed with the noise measuring modes which would invalidate what I've said so far.

Edit:

Well it looks like Takehiro has added at least a few new modes. Don't know if they are used in the new presets or not..

This post has been edited by Dibrom: Jan 30 2003, 00:54

[LordofStars]

Using -X 1,3 works well for me perhaps the mode you suggest would work well for implementations at low bitrates. Such as streaming.

[mithrandir]

My interest in a new mode came from knowing it wouldn't be very hard to add into the existing code. But I have no idea if such a mode would even have the desired improvement in quality (over -X 1 for long blocks)...it was one of those "has anybody thought about this?" things.

Gabriel is using -X 3 for long blocks for some of the APS candidates in LAME 3.94 alpha8/9. APS is not a lower bitrate solution so using -X 3 for long blocks may be fine but from a bitrate control perspective, it's a little "risky". So I thought, why don't we try something that's a little toned down? Of course, I don't have a background in psychoacoustical theory so I'm just brainstorming.

[Gabriel]

X3 is fine when you are targetting high bitrate. X3 is also fine when you are targetting low bitrates, when there is distortion in every sfb.

But when your target is medium bitrate, it's more problematic.

Quantization comparison criterion (-X) is not everything, unfortunately. In order to achieve correct results, it's likely that you will have to change more things than only the X mode. If it would be that easy...